Crypto AI inference undercutting OpenRouter, Platonic representation theorem, payments as the new ads, Google’s $15B Malaysia data centers, Apple scanning iCloud public links
Hello, dear TEA-mates! Here is what you need to know today.
1. ⛓️ Crypto’s Advantage in the Era of Commoditization
In an X article, CLR argues that decentralized networks are undercutting centralized AI inference on price. On the routing layer, Surplus lets holders of excess API credits on platforms such as OpenAI, DeepSeek and Anthropic list discounted inference and get paid in USDC on Base without KYCing every seller, unlike OpenRouter’s manually approved provider list. On the serving layer, Engy (Bittensor subnet 53) has miners compete on optimizations, cheap electricity and tighter margins, which CLR says beats CoreWeave-style operators limited by their own engineering, power costs and investor-driven margins. Over 90 days, Surplus token usage rose from about 10 billion to over 570 billion tokens a day and requests from about 143,000 to 11.2 million a day; Engy rose from 2.7 billion to a peak of over 153 billion tokens a day, with buybacks climbing from about $363 to over $8,000 a day. CLR cites a roughly $160 billion inference market growing toward $350 billion next year, with Surplus at an $18.3 million market cap and Engy at $41.7 million. (Read More)
🫖 TEA For Thought: “This is so cool. I wonder about safety though, even though it’s cheap. That leads us to the dilemma of whether security, decentralization, and scalability can exist at the same time in the crypto world. These networks are built for price sensitivity and scale, not strict corporate compliance. For autonomous coding agents, scrapers, automated pipelines, and non-sensitive background tasks where token volume dominates budget concerns, they offer massive cost savings. For regulated healthcare, financial records, or confidential enterprise IP, traditional direct providers remain the safer choice.”
2. 🧠 Proving and Improving the Platonic Representation Hypothesis with JEPAs
Randall Balestriero and Yann LeCun turn the Platonic Representation Hypothesis into a theorem on arXiv. They prove that Joint-Embedding Predictive Architectures trained separately on any number of modalities learn the same representation up to one rotation when their positive pairs are two observations of one scene (”good views”) and misalignment is bounded otherwise. They show next-token prediction is a JEPA with “bad views” (parts of a single document), which explains why language models align with vision mainly through scale. Validating from 145,000 captions to the texts of 940 million web images, they report a scaling law of alignment and find good views need 42 times fewer training samples than next-token prediction on the same corpus. A 109-million-parameter text JEPA aligned with DINOv2 and CLIP’s image encoders better than every language model they evaluated, up to 72 billion parameters, and better than CLIP’s text encoder; one rotation turned separately trained encoders into an image captioner (CIDEr 0.838 on COCO). (Read More)
🫖 TEA For Thought: “This is so fascinating. Language is only such a small set of intelligence. So much info is missed out in language communication. And what’s fascinating is that when different AI models are trained on different modalities (like text, images, or audio) and scaled up, they naturally converge to learn the exact same underlying representation of the real world. That makes me wonder if that is the representation of how the whole world is like a big simulation that’s grounded in a single rule.”
3. 💳 Payments Are the New Ads
In Fintech Brainfood, Simon Taylor argues that as AI shopping agents grow, platforms need revenue beyond ads, and payments are the emerging monetization layer. Six in ten UK adults told YouGov they would stop using an AI shopping agent after one mistake, while Stripe Link users via Instinct spend about $1,300 a month on average when they stick with it. He maps four card-rail paths today: typing a real card, one-time virtual cards, bank-issued network tokens, and a prefunded agent balance; Instinct’s Balance lets users load a budget (one example adds $700) so the bank sees one top-up and never sees the agent. Merchants are moving faster than banks: Walmart said on October 21 that each customer agent will get its own profile with per-purchase approval or limits such as $50, $100 or $200. Taylor notes Google linked slower Search growth partly to lower monetization in AI Mode, while Google Cloud revenue doubled to $30.7 billion, and Meta is explicitly chasing payments revenue through Muse. (Read More)
🫖 TEA For Thought: “Yeah, the title is a little bit misleading. It should be \”Payments Is a New Monetization Model.\” The agents could be free, but you can charge business owners if they want their agents to purchase the product and if the payments go through. This is going to be the new norm.”
4. 🏢 Google Commits to $15bn Investment in Malaysian Data Centers
Google plans to invest $15 billion to expand its data center infrastructure in Malaysia during 2027 and 2028, Data Center Dynamics reports, citing Bloomberg. The money is aimed at new computing hubs plus water and power infrastructure, focused on Google’s existing campus in Elmina Business Park, Selangor, and new campuses in Selangor and Negeri Sembilan. Google bought 389 acres in Port Dickson, Negeri Sembilan, in May 2025 for a site expected to reach between 500MW and 1GW, with Gamuda Engineering on the works; the existing campus broke ground in October 2024 and its first 100MW phase is due this year. The projects are expected to support more than 9,000 local jobs a year, and Google plans to bring 1GW of clean energy capacity to Malaysia’s grid, after committing $160 million earlier in the week for shared water infrastructure in Negeri Sembilan and $2 billion to the country in 2024. (Read More)
🫖 TEA For Thought: “American politics pushed data centers overseas, which could be a huge national security risk. Those data centers are being built in Asia. Pretty much, Asia is controlled by the CCP, the Chinese Communist Party government. In this case, anything could go wrong. It’s kind of like in sci-fi when your data centers are ruined, when everything is wrong, basically on data centers, this is not good.”
5. 🍎 Apple Will Start Checking iCloud Public Links for Harmful Content
With the first iOS 27.3 beta, Apple published a support document saying that starting in version 27.3 of iOS, iPadOS, macOS and visionOS, items shared with “Anyone with the link” will be checked by Apple for harmful content, 9to5Mac reports. Users can still share with “People you choose” or create a public link; the change applies to the public option for content such as iCloud Photos and iCloud Drive files. If Apple flags shared content as harmful, it “reserves the right to move, refuse, modify, remove, or delete the content at any time,” and may take action against the Apple Account, including blocking or disabling it or involving law enforcement. 9to5Mac notes that public links were already accessible to anyone with the link, including Apple; what is new is that Apple says it will begin checking those links for harmful content. (Read More)
🫖 TEA For Thought: “The question one has to ask is what defines content to be harmful? Who gets to define it? Or is it just a fancy way of scanning your data, controlling the data in the name of security?”
🛠️ Skill of the Day
The Claim Triangulator: Forces any bold claim through three independent angles before you trust it.
You are a careful research analyst who refuses to take a claim at face value.
The claim: [PASTE THE CLAIM HERE]
What I already know: [OPTIONAL CONTEXT]
Sources I can check: [LINKS, NOTES, OR "NONE"]
Step 1: Restate the claim in one plain sentence and name the single measurable fact it depends on.
Step 2: Find or invent three independent angles that could confirm or break that fact: (a) a primary source or official number, (b) a second outlet or dataset that should agree if the claim is true, (c) a falsifier, meaning what evidence would prove the claim wrong.
Step 3: For each angle, say what you found, what is still missing, and how strongly it supports or weakens the claim (strong / mixed / weak / unknown).
Step 4: Write a one-paragraph verdict: trust, trust with caveats, or do not trust yet. Name the one follow-up check that would change your mind fastest.
Rules: Do not invent numbers, quotes, or study titles. If a source is missing, say "not found" instead of guessing. Keep the whole answer under 350 words. No hype language.Paste it into ChatGPT, Claude, or your tool of choice. Replace the bracketed bits with your own.
TEAHEE Moment
Stay sharp, stay informed. See you Monday.
If you enjoyed this TEA, follow along on social for more: